{
  "schema_version": "1",
  "model": {
    "id": "mlx-community/Phi-3.5-MoE-instruct-4bit",
    "revision": "c55613aaad97402ea5500941a0ea458e63714a57",
    "architecture": "phimoe",
    "quantization": {
      "group_size": 64,
      "bits": 4
    },
    "n_experts": 16,
    "top_k": 2,
    "moe_layers": 32
  },
  "strata": {
    "commit": "6562877",
    "mlx_lm": "0.31.3"
  },
  "mode": "streamed",
  "measured_on": {
    "chip": "Apple M5 Max",
    "ram_gb": 137.4,
    "os": "macOS 26.5.1",
    "mlx_lm": "0.31.3"
  },
  "workload": {
    "prompt_set": "bench/mixed-v1",
    "prompt": "Explain how a CPU pipeline works.",
    "generated_tokens": 140,
    "warmup_tokens": 40,
    "decoding": "greedy"
  },
  "memory": {
    "target_ram_gb": 16.0,
    "reserve_gb": 4.0,
    "available_gb": 12.0,
    "enforced_memory_gb": null,
    "full_footprint_gb": 23.55,
    "measured_peak_gb": 10.19,
    "peak_rss_gb": 2.15,
    "swap_growth_gb": 0.0,
    "compression_events": 0,
    "expert_budget": 5,
    "working_set_gb": 7.98,
    "mode": "paged"
  },
  "throughput": {
    "tokens_per_s": 6.81,
    "p50_ms": 142.9,
    "p95_ms": 226.6,
    "ssd_reads_mb_per_tok": 1008.6
  },
  "fidelity": {
    "method": "decode-mode teacher-forced, paged vs resident",
    "prompt_set": "arch_fidelity/mixed-v1",
    "budget": 2,
    "positions": 113,
    "agreement_pct": 100.0,
    "perplexity_delta_pct": 0.0,
    "high_confidence_flips": 0,
    "max_abs_dlogit": 0.0,
    "mean_kl": 0.0,
    "bit_exact": true
  },
  "honesty": {
    "device_real": false,
    "physical_constraint_validated": false,
    "measurement_type": "development-host benchmark",
    "memory_peak_scope": "load-through-generation",
    "throughput_is_optimistic": true
  },
  "sample_output": "A CPU pipeline is a technique used in modern microprocessors to improve their performance and efficiency by allowing multiple instructions to be executed simultaneously at different stages of their execution process. It is analogous to an assembly line in a factory, where different stages of a task are performed concurrently by different workers. In a CPU pipeline, the processor is divided into several stages, each responsible for a specific task in the instruction execution process. These stages are typically organized in a sequential manner, and each stage can work independently on a differe"
}